For more on how the project came to life, the players involved, and the results, see the dedicated article and the Bodoni case study. This article goes deeper into the product research process carried out during the months when the project was born and grew.
Generative AI has multiplied the capacity to produce information at unprecedented speed and scale. At the same time, it has also multiplied the capacity to produce disinformation at unprecedented speed and scale. The World Economic Forum ranked it as the primary short-term threat to civil society in both 2024 and 2025 — two years in a row.
Bodoni uses AI to help journalists counter disinformation. The paradox is obvious, but it is not a good reason to stop. To do it sensibly, however, we could not start with the technology. We first needed to talk to the people who would have to use it: understand how they work, what they fear, where they get lost. What follows is an account of that process.
Quality vs. Quantity
The first fork in the road was methodological. On one side, the temptation of quantitative research: large numbers, structured questionnaires, charts to show in presentations. On the other, the slower but richer path of qualitative research: in-depth interviews, narratives, and contradictions — a volume of stories that cannot fit into a table but restores texture to daily work.
We chose the second route. Not out of disinterest in numbers, but because the central question was not “How many journalists are already using AI?” The real questions were: “What role do they want it to play in their professional practice?”, “Where does it help them, where does it worry them, where does it risk making their work worse?” No perfect statistic would have answered any of those questions.
The research team combined three different perspectives: a senior UX researcher from Google, Eleonora Zucconi; the creative director and designer at Geckosoft, Fabio Santaniello Bruun; and an associate professor of media sociology at the University of Milan, Sergio Splendore, who guided the design and validation of the interview protocols. Three different viewpoints united by one objective: to build a product that made sense first of all for those who, every day, must decide what to cover, how to verify it, and how to put their name on it.
A Mythological Figure That Does Not Exist: The Average User
The composition of the sample was also a deliberate choice. The “average journalist” does not exist, and designing for that figure would have produced something that works for no one.
The sample intentionally sought out extremes: small local publishers and large editorial groups, digitally native outlets and traditional media, journalists with fewer than three years of experience and senior staff with over ten years in newsrooms. The result is a mosaic ranging from the local reporter navigating bars, municipal offices, and WhatsApp groups, to the journalist at a major national outlet immersed in continuous flows of wire agency feeds, reports, and foreign news.
This principle is worth pausing on, because it applies to design in general. Designing for the average user is a comfort zone that rarely produces innovation. Curb cuts, born for wheelchairs, are used by anyone with a trolley. Subtitles, born for the deaf, are now used by anyone watching video in public without audio. If an interface works for those who face the most difficulty or those with the highest demands, it is more robust for everyone. The opposite risk also exists, of course: overloading the interface with niche features that confuse the majority. The key is information architecture: the solution for the extreme case must be integrated, not bolted on as an extra piece.
A Badly Framed Question Distorts the Answer
The interviews, lasting 45 to 60 minutes each, were built with the care of people who understand that the subject at hand requires delicacy.
Under Sergio Splendore’s guidance, an interview guide was developed around three areas. The first concerned daily workflow: how an article takes shape, who decides on topics, how sources are selected, what happens between the initial idea and publication. The second focused on tools: from the CMS to video editing software, from institutional databases to editorial chat channels, to early attempts at using generative AI such as ChatGPT, Gemini, or Copilot. The third addressed the most sensitive area: risks and opportunities of AI in journalism, fears around job loss or loss of credibility, but also curiosity and experiments already underway.
The interviews filled up with cases: an investigation into real estate fraud reconstructed by cross-referencing registries, student chats, and testimony from former employees; a viral video dismantled through reverse image search; a wrong headline that had damaged the perception of an entire piece. It was in these stories that the research found the living material from which to start.
The Journalist’s Map
Once the interviews were collected, the work shifted from listening to interpretation. Transcriptions, thematic coding, comparisons across similar and contrasting contexts: from this effort emerged a shared representation of the journalist’s journey in content creation.
The six-stage map — topic selection, source selection, information gathering, content production, publication, impact monitoring — seems almost obvious at first glance. It was in how it was filled in that it became a powerful tool.
For each stage, the research surfaced deep motivations (the desire to serve the public interest and build a solid reputation) alongside more pragmatic drives tied to visibility and career. Alongside these came the fears: making a public mistake, failing to keep up with innovations, becoming professionally irrelevant, being attacked on social media.
This emotional map carried equal weight to the operational one. The Reuters Institute Digital News Report 2025 documents that only 33% of the public believes journalists regularly check AI outputs before publication. Trust is already fragile, and a product that erodes it further — even inadvertently — is not a useful product.
Choosing Topics
One of the first stages examined was topic selection. Here, power dynamics within the newsroom, the journalist’s personal interests, market logic, and social media logic all intersect.
Story ideas come from multiple directions: assignments from editors-in-chief, proposals from journalists, unexpected events requiring immediate coverage. In between, there is the awareness that a compelling headline can drive clicks but also shift the center of gravity from news value to aesthetic impact; that sometimes a journalist is asked to write in a matter of minutes on a topic they know nothing about; that the most deeply felt stories struggle to find space and adequate compensation.
From this complexity came design questions. How can an AI system help a journalist discover relevant topics in their niche before they become saturated, without turning them into a trend-chaser? How can it support someone who finds themselves covering an unfamiliar topic with thirty minutes’ notice, by offering quality sources and a clear picture without replacing the verification work? And especially for freelancers, how can it surface connections between topics and the outlets most likely to be interested in buying them?
These questions became product concepts: the ability to suggest stories with limited coverage in specific contexts, or a mode that synthesizes a complex topic in journalistic language, with declared sources and prompts for further investigation.
Choosing Sources
If topic selection is the “what,” source selection is the “from whom” and the “how.” Here the research gathered rich material — refined methodologies and survival strategies built over years of experience.
Journalists navigate an archipelago of sources: wire agencies, institutions, public and private archives, academic experts, direct witnesses, social networks, and closed groups. The difficulty lies not just in navigating these islands but in doing so while simultaneously managing three competing needs: understanding which sources are relevant, ensuring they are current, and assessing their reliability at multiple levels. As one interviewee explained:
“Today the hardest part of my job is breaking news that is genuinely original. Because news usually passes through institutional channels — it is very hard to work with primary sources… there is too little time, because sources no longer talk like they used to.”
Some of the strategies that emerged for managing all of this: Excel spreadsheets listing sites to avoid; personal address books of reliable contacts; triangulations using trusted wire agencies as sentinels; systematic verification of images and videos through reverse searches.
The design direction emerged clearly: AI should not replace the journalist’s judgment but make visible the information they already struggle to find today. From this came the idea that the system could propose, starting from a topic, a curated set of potential sources with signals about their nature, their recency, and their history of use in other newsrooms. Alongside this, the ability to show how different sets of sources tell the same story: which narrative clusters exist, which voices remain at the margins, which anomalies deserve deeper scrutiny.
Surviving the Noise
Selecting sources is one thing. Living day-to-day inside the continuous stream of notifications, press releases, phone calls, and chats is another.
For those covering local news, the work remains rooted in territory: word of mouth, physical presence, messages from saved contacts and from unknown numbers, WhatsApp groups. For those working on national news, foreign affairs, or long-form journalism, the picture is different but no less chaotic: wire agency feeds scrolling without pause, social media tips, databases to consult, press releases to sift through. A senior journalist described it without softening:
“There is no time to think anymore. So you move on instinct — and how many things would you do differently, better or even worse, if you stopped, if you went deeper? Young journalists operate mechanically; everything moves so fast that the three-dimensionality of the relationship with the story and the audience is lost.”
The Reuters Institute has documented how “AI slop” sites — content automatically generated to game search engines — are already overwhelming professional journalism. In this context, the design concepts that emerged from the research point in the opposite direction: systems that, starting from an alert, build not a ready-made article, but a possible path — sources to contact, relevant historical data, key questions to ask, suggestions on who has worked on similar topics before.
From Research to Design
A central element of the process was how research was fed back to the product team. We did without a single large final report, preferring weekly updates that alternated between journey maps, journalist quotes, analysis of existing practices, and design direction proposals.
Each update followed a recurring structure: description of the problem, reconstruction of how journalists today handle that specific difficulty, hypothesis about how an AI system could intervene sensibly. This format made it possible to transform insights into feature sketches, use case scenarios, and early prototypes almost in real time.
Insights guided the choice of problems to focus on, and early prototypes raised new research questions: what kind of transparency is genuinely comprehensible in an interface? How far is it useful to show all the steps of an AI model without generating further complexity?
Designing for Trust, Control, and Context
The core design principle was “augment, don’t replace”: not building a system that writes on behalf of journalists, but a tool that places them in a better position to do the work where their judgment is irreplaceable. The interviews surfaced different positions. Some expressed a pragmatic view:
“My philosophy is that if you do work a machine can do, the machine will replace you. But if you go out into the street, do interviews, investigative work… GPT doesn’t go out on the street to conduct an interview.”
Others voiced deeper concerns:
“The most important risk is the banality and mediocrity of the stories we write. I am terrified of a future in which stories are written exclusively by robots… which are unable to translate people’s lives into words.”
Both positions are reasonable; both describe a real risk. Neither resolves the question of how to design a tool that holds both together.
In Bodoni AI, the distinction between what is generated by the system and what is the journalist’s decision is always visible. Functions never lead directly to publication — only to the preparation of material that requires a conscious editorial intervention. The same Reuters Institute report documents that only 12% of the public is comfortable with news produced entirely by AI, but that comfort levels rise significantly when human oversight is visible. Designing for that visibility is not an aesthetic detail.
Some details that rarely make it into product brochures but matter in practice: timestamps always visible, explicit links to original sources, clear indicators of the confidence level of results. These transparency features — traceable sources, declared confidence scores, full audit trail of what the AI produced and why — are the same principles that underpin the governance layer in the underlying Rational AI platform.
Then there is context. The same feature can have a completely different meaning in a small local newsroom with a single journalist on duty versus a large daily with multiple desks and dedicated fact-checking staff. Bodoni AI was designed to adapt to these differences: in some cases as a lightweight, near plug-and-play tool; in others as a component integrable into more complex workflows, capable of connecting with existing CMS platforms and internal systems.
Deciding What to Build: The RICE Method for the MVP
At the point where research insights had generated more ideas than available resources, we had to choose what to include in the MVP. The method used was RICE (Reach, Impact, Confidence, Effort) — a framework that balances potential against reality, keeping everything anchored to the project’s north star: strengthening democracy by countering disinformation and promoting reliable information in the news ecosystem.
RICE prioritized two areas. The first concerned content generated by Bodoni based on reliable sources, with technologies that not only produce text but expose transparent references and support verification. The second concerned access to quality information directly within journalists’ workflows, reducing friction and accelerating access to what matters.
From this prioritization came the three product feature areas: the news graph, the overview of most-discussed topics with its interactive map, and the CoEditor for AI-assisted article writing. The news graph in particular is built on a Knowledge Graph architecture — the same approach used in Rational AI’s Knowledge module — which maps relationships between sources, claims, and stories rather than treating content as isolated documents. This allows the system to follow semantic connections across the information space rather than returning surface-level keyword matches. More details are in the case study.
Development, Prototypes, Testing
The move from concept to code was almost natural. The development team worked on targeted proof-of-concept builds, testing the ability of the models to propose useful sources for given topics, synthesize long content in journalistic language, identify inconsistencies across multiple versions of the same story, group news by similar themes, and extract claims to connect them across different sources. Every prototype was conceived not as an abstract demonstration but as a response to stories already heard.
The cycle closed — and reopened — with usability tests conducted again with journalists whose profiles matched those interviewed in the initial phase. The goal was not to understand whether the interface worked, but to observe how the way of working on news changed with Bodoni’s support.
Research as Permanent Infrastructure
When Bodoni AI launched, the research work did not end — it simply changed form. The journalists’ voices gathered in interviews were joined by usage data: which features are used most, which are ignored, where flows break down. The weekly updates stopped being only a discovery log and became a long memory of the product: an archive of cases, maps, and quotes to return to every time a new direction opens.
The most important lesson of this process is this: UX research is not a preliminary phase that closes with a document. It is an infrastructure that sustains the product over time. In an era when it is easy to fall in love with AI models and forget the people who should use them, Bodoni AI is a deliberate attempt to move in the opposite direction.
Workflows first, features second. Journalists’ fears and motivations first, system architecture second. Stories first, code second.
Want to be one of the first professionals to use it? Go to bodoni.ai and join the waitlist.
Originally published on Geckosoft website, translated with Sonnet 4.6 through Rational AI.